Author(s): Rick Hightower

Originally published on Towards AI.

Part 7: Custom tools let your Claude agent call your code. MCP connects it to the outside world. Subagents let it delegate. Here is how all three fit the same agent without bloating it.

Your agent works until the task gets real, then it cannot reach your API, cannot delegate, and drowns its own context in noise. Here is the fix. The built-in Read, Edit, and Bash tools are a floor, not a ceiling. Three moves grow an agent up: custom tools for your code, MCP for the outside world, and subagents for delegation, each one a way to keep the context window from filling with noise.

This article explains why “capable” agents hit limits once real tasks require reaching APIs, delegating work, and managing context size. It lays out three complementary upgrades: (1) custom tools that let Claude call developer-written functions and return structured results (including guidance on read-only hints to enable parallel, side-effect-free tool runs), (2) MCP to connect the agent to external tool servers you didn’t write (including how to configure stdio vs HTTP/SSE MCP servers, how allowedTools gates tool permissions, and how tool search prevents large tool definitions from crowding the context window), and (3) subagents that delegate focused subtasks in isolated conversations so intermediate context doesn’t pollute the main thread—along with key gotchas around what context and permission modes are inherited. The piece concludes by showing how these moves combine into one workflow for a code-maintenance agent that tests, fixes, re-tests, and then gets an independent review before optionally using MCP-backed capabilities like filing issues or pulling external data.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI